Dual-Branch Underwater-Image Enhancement via Bidirectional Supervision, Contrastive Learning, and Style Transfer

Underwater images frequently suffer from wavelength-dependent color attenuation, low contrast, and detail loss. Although learning-based methods have achieved substantial progress, their performance is often limited by the scarcity and insufficient diversity of paired underwater data. This paper presents a dual-branch underwater-image enhancement network that jointly estimates an enhanced reflectance image and an illumination map. The restoration branch focuses on local structure and color reconstruction, whereas the illumination estimation branch captures global lighting information through window-based self-attention. A feature fusion and exchange module enables bidirectional interaction between the two branches, and a hybrid attention module integrates channel, pixel, and local–global contextual cues. To exploit both paired and unpaired data, the model is optimized using an alternating supervised–unsupervised strategy. The supervised stage combines bidirectional supervision, perceptual, and contrastive losses, while the unsupervised stage introduces degradation-consistency constraints and a contrastive style transfer loss to learn the color distribution of natural images without adversarial training. Experiments on TEST90, TEST60, EUVP, and RUIE demonstrate that the proposed method achieves the highest SSIM on TEST90 and competitive or superior no-reference quality on multiple datasets, particularly in NIQE and MUSIQ. Ablation studies further verify the effectiveness of the feature exchange mechanism, hybrid attention, and joint training strategy.